[ty] prevent "tainted" unions in cycle normalization (#23563)
## Summary
There are some issues with our current cycle handling, particularly with
our implementation of `Type::cycle_normalized`.
To prevent values from oscillating endlessly within a fixed-point
iteration, `cycle_normalized` unions the type of the previous iteration
with the current type. This ensures monotonicity of the type inference
calculation.
However, we've observed that always applying this union can sometimes
produce undesirable results. Therefore, we make some exceptions where
this union isn't applied (see #21909, #21910).
However, these prescriptions are likely not exhaustive, and #22794 also
confirmed the occurrence of "divergent pollution" when narrowing. The
core issue in #22794 is that `Divergent` can inhibit narrowing. The
result type itself does not contain `Divergent`, but `Divergent`
indirectly affects it.
This suggests that determining whether a type contains `Divergent` is
not a perfect indicator of "taintedness".
After carefully examining these cases, I found that the problematic
"low-precision types" only appear in the first few iterations:
`Divergent` itself and the type of the next iteration constructed by
referencing it. Therefore, we shouldn't union these few iterations. I
found that this simple change can remove all of the exceptional handling
for `Divergent` introduced in #21909, #21910, and #22794.
I also found that this change removes the too-many-cycle panic blocking
#22633.
## Test Plan
mdtest updated
---------
Co-authored-by: Carl Meyer <carl@astral.sh>
